Does Executive Compensation Depend on Product Market Structure? Evidence from Shocks to Firm Risk

Transcription

1 Does Executive Compensation Depend on Product Market Structure? Evidence from Shocks to Firm Risk Xiaoyuan Hu University of Maryland January 15, 2017 Abstract This paper shows that, depending on product market structure, firms adjust executive compensation differently in response to shocks to firm risk. Using a natural experiment that increases firm risk due to discoveries of carcinogens, I find that treated firms increase CEO risk-taking incentives to mitigate underinvestment. This result is mainly driven by treated firms in less affected industries, which suggests that firms respond to shocks more strongly when fewer rivals face the same shock, and extends existing work on executive compensation adjustments based on industry-level analyses. Keywords: Executive Compensation, Risk-taking Incentives, Firm Risk, Product Markets Job market paper. Robert H. Smith School of Business, University of Maryland, 3330 Van Munching Hall, College Park, MD xiaoyuanhu@rhsmith.umd.edu. Website: xiaoyuanshirleyhu/. I am grateful to members of my dissertation committee: Vojislav Maksimovic (Chair), Laurent Frésard, Richmond Mathews, Liu Yang, and Ingmar Prucha for their guidance and insightful suggestions. I also thank Gurdip Bakshi, Francesco D Acunto, Michael Faulkender, Albert Kyle, Danmo Lin, Mark Lowenstein, William Mullins, Nagpurnanand Prabhala, Alberto Rossi, Shrihari Santosh, Russell Wermers, my PhD colleagues, seminar participants at the University of Maryland, and conference participants at the 2016 FMA Doctoral Student Consortium for helpful comments. All errors are my own.

2 Does Executive Compensation Depend on Product Market Structure? Evidence from Shocks to Firm Risk January 15, 2017 Abstract This paper shows that, depending on product market structure, firms adjust executive compensation differently in response to shocks to firm risk. Using a natural experiment that increases firm risk due to discoveries of carcinogens, I find that treated firms increase CEO risk-taking incentives to mitigate underinvestment. This result is mainly driven by treated firms in less affected industries, which suggests that firms respond to shocks more strongly when fewer rivals face the same shock, and extends existing work on executive compensation adjustments based on industry-level analyses. Keywords: Executive Compensation, Risk-taking Incentives, Firm Risk, Product Markets

3 Introduction The design of executive compensation receives substantial interest from both practice and academia. The average real value of total CEO pay in S&P 500 firms climbed from $1.1 million in 1970 to $10.9 million in Most of the increase in CEO pay during this period is explained by the growth in stock option compensation (Murphy, 2013). Agency theories suggest that firms use compensation contracts to align managers and shareholders interests. A number of studies recognize that riskaverse, under-diversified managers may underinvest in risky and valuable projects (e.g., Jensen and Meckling, 1976; Smith and Stulz, 1985; Guay, 1999). To mitigate this risk-related agency conflict, firms may provide managers risk-taking incentives by using option compensation, the value of which increases with firms stock return volatility (a measure for firm risk). However, existing evidence on the relationship between firm risk and executive compensation is mixed. In addition, it remains under-explored how firms take into account both their own and other firms risk profiles when choosing compensation policies. In this paper, I examine how firms adjust executive compensation in response to unexpected shocks to their firm risk, and compare firms in two types of industries: more affected industries, in which a larger fraction of firms is affected by a same shock to firm risk, and less affected industries, in which a smaller fraction of firms is affected by the shock. How firms respond to shocks may depend on whether their rival firms face the same shock. Subsequent to an unexpected change in firm risk, affected firms in a more affected industry may not need to adjust executive compensation. Since they account for a majority of firms in the industry, they may be able to increase product prices to reflect any unexpected increase in marginal costs. In other words, product prices may absorb common shocks within an industry. However, affected firms in a less affected industry may need to adjust compensation to remain competitive, because idiosyncratic shocks may not be absorbed in product prices. 1 By taking into account the risk profiles of all firms in an industry, this paper helps to understand within-industry variations in executive compensation policies. It is also important to examine the relationship between firm risk and executive compensation, 1 Despite the fact that firms may mitigate the impact of a domestic shock by trading in foreign markets, to the extent that firms cannot fully diversify away the risk, I expect to find different responses between more affected and less affected industries. In addition, firms with market power may mitigate the impact of the shock by adjusting product prices. Thus, I expect to find a larger difference between more affected and less affected industries when both types of industries are more competitive. 1

4 since the relationship between the two is controversial both theoretically and empirically. A group of theoretical studies suggests that convex payoffs like stock option grants induce managers to take risks, because managers share in the gains but not all the losses (e.g., Jensen and Meckling, 1976; Smith and Stulz, 1985; Edmans and Gabaix, 2011). Another group of studies note that option compensation does not unambiguously lead to more risk-taking, because, apart from increasing convexity, it also increases risk-averse managers exposure to firm risk (e.g., Lambert, Larcker, and Verrecchia, 1991; Carpenter, 2000; Ross, 2004). The empirical literature documents a positive effect of convexity given to managers on managerial risk-taking, but mixed evidence on the effect of option grants. 2 Unlike this literature, my paper focuses on a different question: how firms adjust managerial risk-taking incentives when firm risk changes. Existing evidence on this question is also mixed. Based on an industry-level analysis, Gormley, Matsa, and Milbourn (2013) find that firms reduce managerial risk-taking incentives after firm risk increases. They suggest that firms may want to mitigate managers exposure to firm-specific risk. In contrast, Panousi and Papanikolaou (2012) and De Angelis, Grullon, and Michenaud (2015) show that firms increase managerial risk-taking incentives when firm risk rises. Panousi and Papanikolaou (2012) argue that firms may want to mitigate underinvestment by risk-averse managers. One challenge to the related empirical research is that risk and compensation may be jointly determined. To identify a causal relationship, I exploit sudden increases in firm risk when the National Institutes of Health (NIH) discovers the carcinogenicity of a chemical produced or used by a firm. Under a congressional act 3, the NIH formally identifies and issues a list of carcinogens to the public. Between 1980 and 2014, the NIH updated the list 13 times, resulting in a total of 267 carcinogens. Newly discovered carcinogens attract attention from the public, academia, businesses, and policy-makers. Firms producing or using those carcinogens may face increased risk of litigation regarding issues like workplace injuries, consumer product safety, and environmental pollution. The litigation uncertainty may exist for a couple of years before regulatory agencies make specific decisions, which might then lead to sizable firm value losses. For example, formaldehyde was identified in 2011 as a known carcinogen. On February 22, 2016, when the Centers for Disease Control and Prevention confirmed that the formaldehyde-containing products sold by Lumber Liquidators, a 2 See, for example, Agrawal and Mandelker (1987), DeFusco, Johnson, and Zorn (1990), Guay (1999), Rajgopal and Shevlin (2002), Lewellen (2006), Coles, Daniel, and Naveen (2006), and Low (2009). 3 Section 301(b)(4) of the Public Health Service Act, amended in

5 flooring retailer, can cause cancer, the company s stock price plunged by 23 percent. 4 My sample consists of 7,614 treated and control firm-year observations during the period of within six-year windows around discoveries of carcinogens in 1989, 1991, 2000, 2004, and To identify treated and control firms, I first match the NIH s list of carcinogens with plantlevel information on toxic chemical emissions (including carcinogens and non-carcinogenic toxic chemicals) from the Toxic Release Inventory (TRI) database maintained by the U.S. Environmental Protection Agency (EPA). A plant may emit a certain chemical because it produces the chemical for sales, or uses or processes the chemical to produce other products. 6 I then map the parent companies of TRI-reporting plants to firms listed in Compustat. Around one-fifth of the firms in the Compustat database own TRI-reporting plants. Finally, I restrict my sample to firms with available information on executive compensation from Execucomp and Yermack (1995). My final sample includes 370 unique treated firms, or 601 treatment events. I first verify that the discovery of carcinogens leads to an increase in firm risk. Using a differencein-difference methodology, I show that treated firms experience a 5% increase in firms optionimplied volatility within a 12-month window around the discovery, and a 18% increase in stock return variance within a six-year window. 7 I then examine how firms adjust managerial risktaking incentives, measured mainly by CEO flow vega, the sensitivity of a CEO s current-year compensation to stock return volatility. Treated firms increase the value of CEOs current-year compensation, on average, by $2,859 per 0.01 increase in the firms stock return volatility, which accounts for around 10% of the sample mean of flow vega prior to the discovery of carcinogens. The result is mainly driven by less affected industries, in which treated firms increase flow vega by around 20%. Figure 1 illustrates the different incentive adjustments between more affected and less affected industries. Firms adjust CEO risk-taking incentives more strongly when they are among a smaller fraction of firms in the industry facing the same shock to firm risk. One possible explanation is that product prices may absorb common shocks within an industry, but 4 Source: 5 I focus on these five discoveries of carcinogens due to data availability. 6 I exclude from my control sample firms that never own a TRI-reporting plant during my sample period, because those firms may have a much lower probability of emitting carcinogens and have different characteristics from firms with TRI-reporting plants. See Faulkender and Petersen (2012) for a discussion on related empirical strategies. 7 I find that the average cumulative abnormal returns (CARs) of treated firms is only -0.7% to -0.5% within a three-day or five-day window. One possible reason for the small CARs is a survivorship bias, since the discovery induces some plants to exit. Alternatively, the market reactions could suggest that the discovery mainly affects firm risk rather than expected firm value. 3

6 not idiosyncratic shocks. Treated firms in more affected industries may increase product prices to reflect any increase in marginal costs when they need to switch to new inputs or products. Thus, those firms may not need to react strongly. However, treated firms in less affected industries may need to adjust incentive contracts to remain competitive. Consistent with this explanation, I find that the increases in option-implied and stock return volatility are both driven by treated firms in less affected industries More Affected Eg. Plastics (SIC 2821) % Change in CEO Flow Vega Control Firms Before Treated Firms After Treated Control Less Affected Eg. Cosmetics (SIC 2844) % ChangeTreated in CEO Firms Flow Vega Control Firms Before After Treated Control Figure 1: CEO Incentive Adjustments: More Affected vs. Less Affected This figure presents the percentage change in the average value of CEO incentives (F lowv ega) in a given year relative to the average value of CEO incentives in the year prior to the discovery of carcinogens, by treated and control groups, and by more affected 4-digit SIC industries and less affected 4-digit SIC industries. F lowv ega is defined as the sensitivity of a CEO s current-year compensation (in $000s) to the firm s stock return volatility. The increase in CEO risk-taking incentives is consistent with the argument that firms want to mitigate underinvestment by risk-averse, undiversified executives when idiosyncratic risk rises (Panousi and Papanikolaou, 2012). I provide further evidence that the increase in CEO incentives is accompanied by more R&D expenditures, which is also driven by less affected industries. Also, the increase in CEO incentives is more evident in treated firms that keep producing or using the newly discovered carcinogens than those firms that stop producing or using those carcinogens. In addition, I test several alternative explanations and find no consistent evidence. First, the increase in CEO vega does not seem to be mainly driven by an increase in total pay rather than risk-taking incentives, since treated firms in less affected industries grant more options to their CEOs, but do not significantly change cash or stock compensation, which remains robust after 2005, when the adoption of FAS 123R reduced the accounting advantage of using options. Second, the increase 4

7 in vega is not driven by firms with greater risk-shifting incentives, measured by ex-ante higher leverage or financial distress. Finally, the results are not driven by firms with weaker governance strength, measured by ex-ante lower board independence or lower active institutional ownership. I further explore firm heterogeneity, and show that the increase in CEO incentives is more evident in treated firms producing the newly discovered carcinogens for sales than those firms using the carcinogens to manufacture other products. Furthermore, the increase in incentives is more pronounced in firms with fewer foreign sales or subsidiaries. These results suggest that compensation adjustments depend on how easily a firm can substitute or outsource carcinogen-related production. Consistent with my main findings, less affected industries also drive the subsample results. I conduct several robustness tests. I find no preexisting trends in CEO incentives prior to the shock to firm risk. In addition, my results are robust to controlling for several firm and CEO characteristics and their interactions with the shock. 8 To further account for ex-ante differences between the treated and control groups, I match treated and control firms based on firm characteristics. Furthermore, I find robust results using the number of option grants to proxy for CEO risk-taking incentives, which suggests that the results are not purely driven by mechanical changes in stock volatility. Also, my findings remain similar when I exclude the 2004 discovery of carcinogens, which affected the most number of firms. Finally, I use alternative specifications such as alternative fixed effects, standard errors, and estimation windows. This paper contributes to the literature on the linkage between product markets and managerial compensation. My analysis indicates that, in response to shocks to firm risk, firms adjust managerial compensation more strongly when fewer rival firms face the same shock. A group of studies explore the effect of product market competition on managerial incentives (e.g., Holmstrom, 1982; Hart, 1983; Schmidt, 1997; Raith, 2003). Unlike these studies, I distinguish between product markets in which firms face common shocks to firm risk and those in which few firms face idiosyncratic shocks. Another group of studies investigates how managerial compensation depends on firm performance relative to peers (e.g., Murphy, 1985; Aggarwal and Samwick, 1999). This 8 Controlling for the interaction terms helps to mitigate the concern of bad controls. Simply including a number of control variables could bias the estimation results if the shock to firm risk affects the value of control variables or the way the dependent variable depends on the control variables. See Angrist and Pischke (2009), pp , and Roberts and Whited (2012) for more details on the issue of bad controls. 5

8 paper emphasizes unexpected changes in firm risk rather than relative performance. Also, the evidence in this paper does not seem to suggest that discoveries of carcinogens reveal bad managerial decisions in less affected industries. Furthermore, a strand of literature shows that firms strategically choose peer companies to justify their own compensation policies (e.g., Bizjak, Lemmon, and Naveen, 2008; Faulkender and Yang, 2010). Instead of focusing on self-selected peers, I investigate same-industry firms that face a similar exogenous shock. Finally, the industry equilibrium models of investment and financing decisions suggest that a production technology chosen by many firms in the industry becomes a natural hedge (e.g., Maksimovic and Zechner, 1991; Williams, 1995). Unlike this literature, I examine firms compensation decisions and exploit shocks to firm risk. My findings suggest that firms compensation adjustments are determined by not only their own risk, but also by the risk profiles of other firms in an industry. In addition, this paper contributes to existing work on the effect of firm risk on executive compensation. Building on the industry-level analysis in Gormley, Matsa, and Milbourn (2013), this paper uncovers additional evidence on within-industry variations in compensation adjustments. Gormley, Matsa, and Milbourn (2013) highlight the importance of firm risk in the design of compensation contracts and examine exogenous increases in firm risk due to discoveries of carcinogens. Their study assumes that all firms in more affected industries were treated and all firms in less affected industries were controls due to data availability, and finds that more affected industries reduce CEO flow vega relative to less affected industries after the discovery of carcinogens. 9 I find a similar industry-level result by replicating their empirical strategy based on my sample. In addition, using micro-level data, I show that treated firms increase CEO flow vega, and the increase is driven by treated firms in less affected industries. The results suggest that firms may increase managerial risk-taking incentives in order to mitigate underinvestment by risk-averse, under-diversified managers. The rest of this paper proceeds as follows. Section 1 reviews the related literature and develops the hypotheses. Section 2 discusses the data. Section 3 presents the results and robustness checks. Section 4 concludes. 9 Gormley, Matsa, and Milbourn (2013) use the industry-level National Occupational Exposure Survey to identify treated and control groups. 6

9 1 Related Literature and Hypotheses Development 1.1 Industry Equilibrium of Corporate Decisions A body of theoretical literature explores the industry equilibrium of investment and financing choices (e.g., Maksimovic and Zechner, 1991; Williams, 1995; Fulghieri and Suominen, 2012). Maksimovic and Zechner (1991) demonstrate that the risk of a firm s technology choice is endogenously determined and depends on the equilibrium number of firms choosing each type of technology in an industry. In their paper, each firm can invest in either a technology with a known marginal cost or another technology with an uncertain marginal cost. Suppose that a given firm faces an unexpectedly high cost. If most other firms in the industry have chosen the same production technology and therefore experience the same shock in production costs, then the higher production costs will translate into a higher price of the products sold. Thus, that technology becomes a better hedge and will generate less risky cash flows. In contrast, if a firm is one of few firms in the industry facing an unexpected increase in costs, prices will not reflect the higher production costs. Hence, that technology will exhibit riskier cash flows than the other one. Unlike their study, this paper focuses on compensation decisions rather than investment and financing policies, and investigates unexpected shocks to firm risk rather than endogenous choices of technologies. Nevertheless, the theoretical intuition in their study can be extended to develop my hypothesis. In this paper, I explore how firms adjust executive compensation following shocks that increase firm risk. Specifically, I examine firms responses to the discovery of carcinogenicity of a chemical produced or used by the firms. Because of potential litigation and reputation concerns, affected firms may switch to non-carcinogenic or less toxic chemicals, which may result in higher marginal costs of production. If most of the rival firms in the industry face the same shock to firm risk, affected firms may be able to increase product prices to reflect any increase in marginal costs. In contrast, if only a small fraction of firms in the industry experiences the shock, these firms may not be able to affect product prices. Hence, I expect that treated firms in less affected industries would experience a larger increase in firm risk, and thus would be more responsive to the shock in adjusting executive compensation. This leads to my first hypothesis: Hypothesis 1. If a firm is among a small fraction of companies in the industry facing the same 7

10 shock to their firm risk, the firm would more actively adjust executive compensation, compared to the case in which most of the firms in the industry face a same shock to their firm risk. One implicit assumption of this hypothesis is that firms cannot fully diversify away the increased risk through international trade. In my setting, treated firms might be able to mitigate the impact of the discovery of carcinogens by outsourcing their carcinogen-related production to countries with less rigid regulations than the United States. Thus, I expect the above hypothesis to hold to the extent that firms cannot fully diversify away the risk. 10 In addition, treated firms with market power may be able to adjust product prices even if they are in a less affected industry. Hence, the hypothesis is more likely to hold in competitive markets in which firms take product prices as given. This paper is also related to literature on the relationship between product market competition and corporate outcomes. A strand of studies explores the effect of competition on managerial incentives. For instance, Holmstrom (1982), Hart (1983), and Hermalin (1994) argue that competition reveals additional information about managerial ability if firms in a product market are hit by common productivity shocks. Hermalin (1994) shows that the best response to other firms providing weak (strong) managerial incentives can be to provide strong (weak) incentives. Schmidt (1997) and Raith (2003) suggest that competition drives down firm profits, and thus may discourage managerial efforts but may also discipline managers and boost productivity. Unlike these studies, this paper distinguishes between product markets in which firms face common shocks to firm risk and those in which few firms face idiosyncratic shocks. In another related study, Hadlock and Sonti (2012) examine how revisions to firms asbestos liabilities affect market reactions to their competitors. Unlike their paper, I exploit exogenous shocks to firm risk rather than self-reported revisions to litigation liabilities and focus on treated firms rather than their competitors. In addition, this paper is related to the literature on relative performance evaluation and compensation peer benchmarking. A strand of literature explores whether managerial compensation is determined by firm performance relative to peers (e.g., Murphy, 1985; Aggarwal and Samwick, 1999). This paper examines unexpected changes in firm risk rather than relative performance, and finds no evidence that discoveries of carcinogens reveal bad managerial decisions in less affected industries. Other studies document that firms strategically choose peer companies that pay 10 This may occur, for example, due to uncertainties about foreign trade, imperfect foreign product markets, etc. 8

11 higher executive compensation to justify their own compensation policies (e.g., Bizjak, Lemmon, and Naveen, 2008; Faulkender and Yang, 2010; Albuquerque, De Franco, and Verdi, 2013). Instead of focusing on self-selected peers, I investigate same-industry firms that face a similar exogenous shock to their firm risk. 1.2 Firm Risk and Executive Compensation Existing literature provides mixed evidence on the effect of firm risk on compensation decisions. Using industry-level evidence, Gormley, Matsa, and Milbourn (2013) investigate how firms adjust managerial incentives when their liability risk rises due to workplace exposure to newly discovered carcinogens. Their study assumes that all firms in more affected industries were treated, and all firms in less affected industries were controls, and shows that more affected industries reduce CEO flow vega relative to less affected industries after the discovery of carcinogens. 11 This paper extends their work by using micro-level data and exploring within-industry variations in compensation adjustments. Another strand of studies suggests a positive effect of firm risk on managerial risk-taking incentives. For instance, De Angelis, Grullon, and Michenaud (2015) investigate a sudden increase in downside firm risk following removal of short-selling constraints, and show that treated firms grant relatively more stock options to their executives than restricted stocks. A recent study by Panousi and Papanikolaou (2012) distinguishes between systematic and idiosyncratic components of risk, and argue that top executives can hedge away their exposure to systematic risk but not to idiosyncratic risk, since they are not permitted to buy put options or short their own company s stock. They use a theoretical model to show that risk-averse, undiversified managers may underinvest in projects characterized by idiosyncratic risk, and risk-neutral, well-diversified shareholders may want to increase managerial risk-taking incentives to mitigate the underinvestment. 12 Consistent with their paper, existing studies on option repricing also suggest that the wedge between managers and shareholders optimal decisions increases with idiosyncratic risk, managerial risk aversion, and the extent of under-diversification (e.g., Hall and Murphy, 2000; Chidambaran and Prabhala, 2003; 11 The treated group in Gormley, Matsa, and Milbourn (2013) consists of a set of SIC industries in which above a threshold fraction of workers is exposed to newly discovered carcinogens. 12 Armstrong and Vashishtha (2012) also assume that managers can hedge against systematic risk rather than idiosyncratic risk and find that vega encourages managers to increase systematic risk more than idiosyncratic risk. Unlike their paper, I focus on a sudden increase in idiosyncractic risk. 9

12 Ingersoll Jr., 2006). 13 The intuition on the relationship between idiosyncratic risk and managerial incentives can be extended to my paper. Under the assumption that managers are risk-averse and under-diversified and cannot flexibly sell or hedge against idiosyncratic risk, prior studies suggest two competing hypotheses on the effect of idiosyncratic risk on managerial incentives: Hypothesis 2a. Firms would reduce managerial risk-taking incentives if idiosyncratic risk increases. Hypothesis 2b. Firms would increase managerial risk-taking incentives if idiosyncratic risk increases. The two competing hypotheses have different implications from an agency theory perspective. Hypothesis 2a implies that firms cut managerial incentives to meet managers constraints, while Hypothesis 2b suggests that firms increase incentives to maximize shareholders value. Following the theoretical framework of Holmstrom and Milgrom (1987), I consider a firm that designs incentive compensation contracts to maximize shareholders value subject to managers participation constraint. Since the manager is risk-averse, a sudden increase in idiosyncratic risk may raise her marginal costs of exerting efforts. Hypothesis 2a suggests that the firm would cut managerial incentives to mitigate the increase in the costs of efforts and meet managers constraints. The finding of Gormley, Matsa, and Milbourn (2013) is consistent with this hypothesis. In contrast, Hypothesis 2b, consistent with Panousi and Papanikolaou (2012), argues that the firm would give managers a greater reward of risk-taking in order to maximize shareholders value. Another body of literature examines the impact of option compensation on managerial risktaking activities. In general, this literature suggests that vega (sensitivity of a manager s wealth to firm volatility) induces managerial risk-taking, while the effect of delta (sensitivity of a manager s wealth to changes in stock price; also known as pay-for-performance sensitivity) or option grants depends on managerial risk aversion, ability to hedge, and outside wealth. 14 A number of theoretical studies demonstrate that option compensation increase vega, which encourages managers to take risks, because the expected payoff of an option increases in the volatility of the underlying stock s 13 These studies show that repricers tend to be smaller, younger firms that experienced an abrupt decline in growth and profitability. In contrast, my sample consists of larger, older firms. 14 See also Armstrong and Vashishtha (2012) for a discussion on this body of literature. 10

13 return (e.g., Jensen and Meckling, 1976; Myers, 1977; Haugen and Senbet, 1981; Smith and Stulz, 1985; Edmans and Gabaix, 2011). Other studies argue that besides vega, option grants also increase delta, which makes managers firm-specific wealth more sensitive to changes in stock prices (e.g., Lambert, Larcker, and Verrecchia, 1991; Carpenter, 2000; Ross, 2004). Thus, if managers are riskaverse and cannot sell or hedge against the risk associated with their options, they may be less willing to take risks. Consistent with the theoretical predictions, empirical studies show that higher vega is associated with greater managerial risk-taking activities, measured by higher stock return volatility, more R&D investments, lower capital expenditures, and higher leverage (e.g., Coles, Daniel, and Naveen, 2006; Low, 2009). There is mixed evidence on the effect of delta or option grants on managerial risk-taking (e.g., Agrawal and Mandelker, 1987; DeFusco, Johnson, and Zorn, 1990; Guay, 1999; Rajgopal and Shevlin, 2002), which suggests that the effect may depend on empirical values of managers risk aversion, wealth, and hedging ability. Unlike this literature, my paper investigates how firms adjust executive compensation in response to changes in firm risk, rather than how firms use compensation contracts to induce managerial risk-taking activities. 2 Data I rely on several data sources to construct my sample. First, I collect the timing of discoveries of carcinogens from the Report on Carcinogens (RoC) prepared by the National Institutes of Health (NIH). Second, to identify treated and control groups in the event of discoveries, I match the RoC data with the plant-level Toxic Release Inventory (TRI) data from the U.S. Environmental Protection Agency (EPA), which differs from existing literature. The TRI data contains plantlevel annual information on toxic chemical usage and emissions (including carcinogens and noncarcinogenic toxic chemicals). Third, I use the Compustat database to obtain firm-level financial data. I aggregate the plant-level chemical data to the firm level and match the parent companies of TRI-reporting plants to firms listed in Compustat. Finally, I focus on a sample with available information on executive compensation from the Execucomp database and Yermack (1995). I focus on the discoveries of carcinogens in 1989, 1991, 2000, 2004, and 2011, due to data availability. Following Gormley, Matsa, and Milbourn (2013), I construct a pooled sample, which consists of five cohorts of treated and control firm-year observations. Each cohort is a six-year 11

14 period around the discovery. In robustness tests, I use alternative estimation windows. Next, I discuss the data and sample construction in details, and present summary statistics. 2.1 Discoveries of Carcinogens In this paper, I exploit discoveries of carcinogens by the NIH as exogenous variations in firm risk. Data on the timing of discoveries is available through the RoC, which is a congressionally mandated, science-based, public health document prepared by the National Toxicology Program of the NIH. 15 Section 301(b)(4) of the Public Health Service Act, amended in 1978, requires that the NIH publishes and updates a list of chemicals, either known to be or reasonably anticipated to be human carcinogens. The RoC provides important information that supports decision-making by the public, businesses, and regulatory agencies. 16 Between 1980 and 2014, the NIH updated the RoC 13 times, resulting in a total of 267 listed carcinogens. Table A.2 in the appendix reports the years of discoveries (column 1). On average, each edition of the RoC includes around 20 newly discovered carcinogens (column 3) and affects 185 Compustat firms (column 5). Among all discoveries, the 2004 discovery affected the most firms (643 firms), followed by the 1989 discovery (312 firms) and the 2000 discovery (245 firms). I obtain the announcement dates of the RoC by searching the news articles published by the National Institute of Environmental Health Sciences (NIEHS) and the National Center for Biotechnology Information (NCBI). 17 The dates are available since 1994 and listed in column I exclude all delisted chemicals and firms affected by delisting. The number of newly discovered carcinogens presented in Table A.2 already excludes delisted chemicals. Delisting is not common. Between 1980 and 2014, only nine chemicals were once discovered as carcinogens but then delisted later (column 4). Only 54 Compustat firms in 2000 were affected by two delisted chemicals (column 6). The reasons for delisting include a low possibility of human exposure and insufficient evidence 15 Available at 16 The regulatory agencies that cite the RoC include Centers for Disease Control and Prevention (CDCP), EPA, Occupational Safety and Health Administration (OSHA), Consumer Product Safety Commission (CPSC), and the Food and Drug Administration (FDA). 17 Sources: The NIH scheduled the 2004 discovery announcement in that year but actually released it to the public in January My findings could be biased for this particular year if firms adjusted executive compensation prior to the actual announcement based on leaked information. However, in robustness tests, I exclude the 2004 discovery and find similar and even stronger results for other years of discoveries. Thus, the inclusion of the 2004 discovery only works against finding the results in my paper. 12

15 of carcinogenicity after reevaluation. In each edition of the RoC, the NIH indicates whether a chemical is known to be or reasonably anticipated to be a human carcinogen. I identify newly discovered carcinogens via their first appearance on the RoC, whether or not they are known or reasonably anticipated to be carcinogens. For instance, formaldehyde was listed as a reasonably anticipated carcinogen in 1981, and then updated to be a known carcinogen in I treat formaldehyde as discovered in For robustness checks, I identify newly discovered carcinogens via their first appearance as known carcinogens, that is, I treat formaldehyde as discovered in Alternatively, I examine a subsample of chemicals that first appeared on the RoC as known carcinogens, that is, I exclude formaldehyde. 2.2 Toxic Chemical Emissions A key step in my analysis is to identify treated and control firms around discoveries of carcinogens. Unlike existing literature, I use a plant-level panel database on toxic chemical usage and emissions (including carcinogens and non-carcinogenic chemicals). I match the RoC data with the TRI data from the EPA. 19 TRI data provides plant characteristics, including plant name, industry, location, chemical characteristics, such as the name of chemicals emitted by a plant and how a chemical is used, and parent company name. A firm may emit a toxic chemical because it produces the chemical for sales or distribution purpose. Alternatively, the firm may use or process the chemical as an input during its production. TRI data has been an important resource for regulators, investors, environmentalists, and communities to assess plant-level and firm-level environmental performance. In 1986, Congress passed the Emergency Planning and Community Right-to-Know Act (EPCRA) to inform the public about toxic chemical emissions in the local community. Under the requirements of EPCRA, all U.S. plants that meet the following reporting criteria must submit annual TRI data to the EPA: (i) The plant is in a specific NAICS industry sector, including manufacturing, mining, utility, wholesalers, etc. 20, or is owned or operated by federal government; (ii) the plant employs 10 or more full-time equivalent employees; and (iii) the plant produces, processes, or otherwise uses one of the TRI-listed toxic 19 Available at 20 Specifically, TRI-covered plant-level NAICS industries include mining (NAICS 212), utilities (221), manufacturing (31 33), miscellaneous manufacturing (1119, 1131, 2111, 4883, 5417, 8114), merchant wholesalers and non-durable goods (424), wholesale electronic markets and agents brokers (425), publishing (511, 512, 519), and hazardous waste (562). 13

16 chemicals in quantities above threshold levels in a given year. I match the TRI data with the RoC data using chemical CAS regristry number and chemical names. 21 Between 1987 and 2014, TRI-reporting plants emitted 610 unique toxic chemicals, including 126 carcinogens and 484 non-carcinogenic toxic chemicals. I aggregate the plant-level chemical data to the firm level, and identify parent companies of TRI-reporting plants listed in Compustat (hereafter refered to as TRI-Compustat firms ). Specifically, I map the name of a parent company in the TRI data to a firm name in Compsutat using a string matching algorithm, then manually check each potential match to improve accuracy. I exclude parent companies with no match from my sample, including potential unmatched Compustat firms and private firms not listed in Compustat. During the period , there are 58,076 unique TRI-reporting plants (656,592 plant-year observations). Around 80% of these observations have available information on parent companies. There are 11,357 unique parent companies with TRI-reporting plants (160,815 firm-year observations), among which 2,415 are identified as Compustat firms (39,939 firm-year observations). Since 1987, discoveries of carcinogens have affected 3,718 (28.0%) out of the 11,357 TRI firms and 878 (36.4%) out of the 2,415 TRI-Compustat firms. The median TRI-Compustat firm has more TRIreporting plants than other TRI firms but emits a lower fraction of carcinogens. Specifically, the median TRI-Compustat firm has 15 TRI-reporting plants, among which 10 plants (66.7%) emit carcinogens. The median TRI-reporting plant of a Compsutat firm emits eight toxic chemicals, among which one chemical (12.5%) is carcinogenic. In comparison, the median TRI firm has six TRIreporting plants, among which five plants (83.3%) emit carcinogens. The median TRI-reporting plant emits six toxic chemicals, among which one chemical (16.7%) is carcinogenic. TRI-Compustat firms account for around one-fifth of the full Compustat database. Compared to other Compustat firms, TRI-Compustat firms are, on average, two times larger and older, grow slower, have better operating performance, and spend less in R&D. 22 My final sample consists of firms with available financial and compensation data, and thus my results mostly apply to large public firms. 21 The Chemical Abstracts Service (CAS) has assigned a unqiue numerical identifier to every chemical described in the open scientific literature since In comparison, the Compustat firms in the S&P 1500 with available compensation data from Execucomp are, on average, four times larger in size than other Compsutat firms. 14

17 Around 90% of the TRI-reporting plants belong to the manufacturing sector (NAICS 31 33). Figure 2 presents the distribution of TRI-reporting plants by the Fama-French 48 industry. The plants are concentrated in a few industries, such as construction materials, chemicals, steel works, rubber and plastic products, and fabricated products. Six Fama-French industries, including real estate, banking, entertainment, insurance, communication, and trading, have no TRI-reporting plant between 1987 and Figure 2 also distinguishes between plants that emit existing carcinogens (black bars) and plants that emit non-carcinogenic toxic chemicals (white bars). Around 40% of the plants emit carcinogens. Compared to the plant-level distribution, the industry distribution of TRI-Compustat firms is less concentrated, as shown in Figure Electronic equipment, machinery, chemicals, and construction materials have more TRI firms than other Fama-French industries. However, they are not necessarily the industries with higher proportions of firms releasing carcinogens. Aircraft industry has the highest fraction of firms releasing carcinogens (84.6%). In 21 Fama-French industries, including food products, computer software, rubber and plastic products, and electronic equipment, less than 50% of the TRI firms emit existing carcinogens (blue color). In the remaining 27 industries, including aircraft, utilities, chemicals, and petroleum and natural gas, over 50% of the TRI firms emit carcinogens (black color). Figure 3 indicates that there are within-industry variations in carcinogen emissions. Thus, the TRI data provides an opportunity to examine the difference between treated firms in more affected industries and treated firms in less affected industries. The use of the TRI data may introduce measurement errors if plants misreport their toxic chemical emissions, but this concern is mitigated in my paper, since I examine exogenous increases in the risk of releasing certain chemicals. Furthermore, the EPA s enforcement policies give plants an incentive to accurately report toxic chemical emissions. Under EPCRA, the EPA conducts compliance inspections, investigates cases of non-compliance, and can issue a maximum civil penalty of $25,000 per violation for not reporting or misreporting emissions One reseaon is that TRI-Compustat firms are more diversified. Compared to the median Compustat firm, the median TRI-Compustat firm has one more segment or unique segment-level 4-digit SIC code. 24 In 2001, the EPA conducted 321 compliance inspections for TRI reporting. Between 1990 and 1999, the EPA brought 2,309 administrative actions against non-compliance (De Marchi and Hamilton, 2006). 15

18 2.3 Sample Construction Following Gormley, Matsa, and Milbourn (2013), I construct a pooled sample, which consists of five cohorts of treated and control firm-year observations. Each cohort is a six-year period (from year T-3 to year T+2) around the discovery of carcinogens (in year T, where T=1989, 1991, 2000, 2004, or 2011). For instance, Air Products & Chemical Inc. (APC) was affected by the 1989 discovery and is included in my sample for the six-year period, between 1986 and I use alternative windows around the discovery in robustness tests. I include firms affected for the second or more times during my sample period and keep overlapping years across cohorts. For example, APC was affected again in 1991, and is included in my sample for the six-year period of 1988 to 1991 for the 1991 cohort, in which overlap with the four years in the 1989 cohort. For robustness checks, I exclude overlapping years, and, alternatively, split the sample between firms affected for the first time and for the second or more times. I focus on the discoveries in 1989, 1991, 2000, 2004, and 2011, because they have available TRI data and because the first four years overlap with the years in Gormley, Matsa, and Milbourn (2013). Restricting my sample to these five years only excludes 7% of treated firms. In robustness tests, I include other years of discoveries. My final sample is restricted to firms with available data on CEO compensation from Execucomp and Yermack (1995). I identify 370 unique treated firms, in total 601 treatment events, since some firms were affected for multiple times. A firm is identified as treated if it owns at least one plant that produces or uses and thus, emits a chemical newly discovered as a carcinogen. Among the 601 events, 231 occurs in less affected 4-digit SIC industries, in which less than or equal to 50% of the TRI firms are simultaneously affected (e.g., cosmetics manufacturing in Figure 4); the remaining 370 events occur in more affected 4-digit SIC industries, in which over 50% of the TRI firms are simultaneously affected (e.g., plastics manufacturing in Figure 4). The pooled sample between 1986 and 2013 has 2,838 treated firm-year observations. My control sample consists of firms with plants that emit non-carcinogenic toxic chemicals. In other words, I exclude firms that never had a plant that emitted any toxic chemical above a threshold value. The rationale is that those firms may have a much lower probability of being affected than TRI firms. In addition, they differ in other firm characteristics from TRI firms. This empirical 25 I do not require a firm in my sample to survive for all six years in a given cohort. 16

19 strategy follows Faulkender and Petersen (2012), who show that different empirical strategies can lead to different estimation results when there is a third group besides the traditional treated and control groups, and combining the third group with the control group may bias estimation results. I identify 624 unique control TRI firms with available compensation data. The pooled sample between 1986 and 2013 has 4,776 control firm-year observations. For robustness checks, I match treated and control firms based on firm size, firm age, and SIC industry. The matched sample consists of 218 unique treated firms (1,158 firm-year observations) and 411 unique control firms (1,683 firm-year observations). 2.4 CEO Compensation I collect executive compensation data from Execucomp and Yermack (1995). Execucomp covers active and inactive firms in the S&P 1500 index since Yermack s sample consists of 792 U.S. firms (5,955 firm-year observations) that appeared in the Forbes magazine lists of the 500 largest public U.S. corporations between 1984 and My main measure for managerial risk-taking incentives is F lowv ega, the change in a CEO s compensation (effectively the value of option grants) during a given year for a 0.01 increase in a firm s stock return volatility. 27 I use the Black and Scholes (1973) formula to value options following Core and Guay (2002) and account for the 2006 change in reporting format in Execucomp following Coles, Daniel, and Naveen (2013). To proxy for risk-free rates, I use the Treasury rate corresponding to the actual option maturity if the option maturity is less than or equal to 10 years, and use the 10-year Treasury rate if the option maturity exceeds 10 years. Stock return volatility is calculated as the annualized standard deviation of monthly stock returns during the past 60 months. To examine whether firms adjust other aspects of compensation besides risk-taking incentives, I collect information on individual components of compensation, including the values of option grants (Options), restricted stock grants (StockComp), salary and bonus (CashComp), and total compensation (T otalcomp, the sum of option, stock, and cash compensation). I use logged values of these measures to mitigate the concern that CEO compensation has a skewed distribution. For 26 Yermack s sample provides information on CEO age, tenure, stock ownership, cash compensation, and option grants based on firms proxy statements, 10-K, and 8-K filings. 27 F lowv ega effectively only accounts for the value of option compensation, since the value of stock or cash compensation do not change with stock return volatility. 17

20 robustness checks, I adopt alternative measures for managerial incentives, including the number of option grants (Options N), flow delta (F lowdelta, the change in the value of a CEO s compensation during a given year for a 1% increase in a firm s stock prices), and vega calculated using current-year plus previous years option compensation. I use the number of option grants to test whether there is a real effect rather than a mechanical effect driven by changes in stock prices and volatility. I use vega to account for a total wealth effect. 2.5 Other Variables Before preceding to my main analysis, I test how the discovery of carcinogens affects firm risk, measured by option-implied or stock return variance. I obtain data on implied volatility from OptionMetrics, and focus on at-the-money call options with at least 90 days to expiration (following DeFusco, Johnson, and Zorn, 1990). I take the open-interest-weighted average value of implied volatility for each firm. The tests are based on a reduced sample, because OptionMetrics is available since Stock return variance is calculated using data from CRSP. I examine whether weakly governed firms drive the results and use the fraction of independent directors on boards (Independence) and potentially active institutional ownership (ActiveOwnership, following Almazan, Hartzell, and Starks, 2005) to proxy for governance strength. The first test is based on a reduced sample, because board information from the ISS (formerly RiskMetrics) database is available since In addition, I examine risk-shifting as an alternative explanation for my results. Risk-shifting incentives are measured by leverage ratio (Leverage) and an indicator for distress based on the Altman (1968) Z-score (Z). In robustness tests, I control for firm and CEO characteristics that may affect a firm s design of CEO compensation. Following Guay (1999), I include firm size (F irmsize), CEO tenure (CEOT enure), and CEO cash compensation (CashComp) as controls. To address the concern of potential bad controls, I interact these controls with the shock to firm risk. In addition, I also control for cash flows, leverage, and CEO age. Detailed definitions of all variables are included in the appendix (Table A.1). 18